Description
Deep-fake detection of voice calls should ideally run on local devices with limited computational, storage, and energy footprint. It has been shown that DNNs specialized for this task perform almost the same when their weights are quantized to ternary values (-1, 0, 1). As a by-product, the resulting weight matrices are extremely sparse. This raises the possibility of compiling the trained network into efficient machine code for sequential execution on commodity CPUs and SoCs.
The objective of this thesis is to develop, evaluate, and document strategies to compile pre-trained and quantized anti-voice spoofing DNNs provided to the student to efficient LLVM intermediate representations. The compilation should be automated and tested for a number of DNN instances. The evaluation should consider runtime and memory use. If further optimizations are possible that alter the structure or weights of the DNN, for example as a result of profiling, the impact on the task performance should be evaluated as well.